Does this model actually get it right?
The model calculates win, draw and loss probabilities. Here we compare neutral forecasts logged before kick-off with actual outcomes. The sample includes available, valid settled forecasts. This evaluates the base match forecast; it does not by itself validate how the model measures the effect of your lineup and tactical changes.
How often the model is right
The page recalculates its metrics from the settled predictions available in the database on every visit.
338 neutral base predictions settled, from 28/08/2026 to 16/10/2026. The window covers up to the 3,000 most recent predictions: only model v0.4.3, season 2026/27 and supported club competitions. Every prediction was logged before kickoff.
Correct 1X2 pick
55 out of 100
55.2%
First-pick accuracy on 317 matches finished in regulation time whose forecast permits a draw. Draws count. Cup forecasts requiring a winner are excluded.
Same-match checks
Forecast coverage
Domestic leagues: 235 / 300 finished matches in the database (78.3%).
338 evaluated forecasts / 746 finished matches in the database (45.3%); 408 without a usable forecast. Window: 2026-08-28 to 2026-10-16. This covers supported competitions in the recorded window, not the provider's full fixture calendar. Missing forecasts can affect the composition of the sample.
Coverage by competition
- Champions League: 36 / 36
- Europa League: 28 / 36
- Premier League: 26 / 50
- FA Cup: 0 / 283
- EFL Cup: 15 / 16
- Ligue 1: 44 / 46
- Bundesliga: 41 / 45
- DFB-Pokal: 1 / 2
- Ekstraklasa: 42 / 50
- Polish Cup: 0 / 27
- Serie A: 35 / 51
- Coppa Italia: 5 / 8
- La Liga: 47 / 58
- Copa del Rey: 0 / 20
- Conference League: 18 / 18
Regulation-time 1X2
175 correct picks / 317 matches (55.2%). Always picking the home side on these same matches: 48.3%. Only matches finished in regulation time whose forecast contract permits a draw; cup forecasts requiring a winner are excluded.
95% Wilson accuracy interval: 49.7%-60.6%. This expresses sample uncertainty, not guaranteed future performance.
Draw probabilities
On the same 317 regulation-time matches, mean predicted draw probability: 21.9%; observed draw frequency: 18.9%. The model picked a draw first in 0 matches. A draw need not be the first pick to receive a meaningful probability.
Brier and RPS: lower is better. Market probabilities are stored implied probabilities after removing the margin. Snapshot timing matters: earlier odds may contain less information; the latest stored pre-kick-off snapshot is not necessarily the closing price.
Odds available before the model forecast
317 identical matches. Median odds snapshot age: 0.5 hours.
| Measure | Dobitka | Odds market |
|---|---|---|
| 1X2 accuracy | 55.2% | 59.6% |
| Brier | 0.563 | 0.532 |
| RPS | 0.205 | 0.190 |
Brier difference (model minus market): 0.031; 95% approximate interval: 0.007 to 0.055. Negative favours the model. The market has a lower Brier score in this paired sample. The approximate interval assumes independent matches; it does not account for repeated clubs. Treat it as evidence from this sample, not proof of lasting superiority.
Latest recorded odds before kick-off
317 identical matches. Median odds snapshot age: 1.0 hours.
| Measure | Dobitka | Odds market |
|---|---|---|
| 1X2 accuracy | 55.2% | 59.6% |
| Brier | 0.563 | 0.532 |
| RPS | 0.205 | 0.190 |
Brier difference (model minus market): 0.031; 95% approximate interval: 0.007 to 0.055. Negative favours the model. The market has a lower Brier score in this paired sample. The approximate interval assumes independent matches; it does not account for repeated clubs. Treat it as evidence from this sample, not proof of lasting superiority.
For the curious: Brier, RPS and calibration
Hit rate alone is a blunt measure - it rewards safe favourites and cannot see whether the probabilities themselves are honest. These metrics check that.
Brier / RPS
0.541 / 0.199
These scores evaluate the full win, draw and loss probability distribution. Lower = better. The fixed reference assigns every match 46% home, 24% draw and 30% away. Its Brier / RPS on the same sample: 0.624 / 0.237. This is a simple baseline, not bookmaker odds.
First-pick confidence calibration
0.062
We group matches by confidence in their most likely outcome and compare each group's mean confidence with its hit rate. MACE: 0.062, the mean absolute gap across non-empty groups, with equal weight per group. Closer to zero = better. This does not separately evaluate calibration of all three probabilities.
Correct pick - mixed sample
57%
A supporting measure with mixed targets: draws count, but cup ties requiring a winner use the eventual winner. This differs from the regulation-time 1X2 sample.
Winner called (278 matches)
69%
69 out of 100
A supporting measure: first-pick accuracy after excluding matches that ended in a draw. The subset is selected after knowing the result, so this is neither all-match accuracy nor a comparison with the 1X2 market.
How we measure it
This sample contains only raw, neutral base predictions from the current model for the club season and supported competitions. The benchmark uses the predicted home XI, default playing style and both teams' strength; it does not contain a specific user's decisions. A prediction must exist before kickoff. After the match we compare the full 1X2 vector with the real result; in cup ties settled after a draw, the winner counts. An additional calibration layer is not applied to the current season's 1X2 result, points or ranking.
We improve the model through separate, versioned releases. Experimental variants do not change points or the ranking. Every change to the result path requires tests, an explicit model version and a human decision - the engine never switches automatically.
Latest settled match in this sample: 16/10/2026.
What the model takes into account
- -The strength of both teams and their player profiles - from current squads and availability. The opponent's XI is added when it can be known before kick-off.
- -Your plan: lineup, formation, playing style and planned substitutions. Each of these shifts the result distribution - otherwise there would be nothing to play for.
- -Each of the eleven players separately: their xG and xA (shot and pass quality), their form over their own last five matches, and the position they actually play. We count player form, not club form.
- -We run the same match a second time without your decisions, on the default setup. That is how you know how much you moved yourself, and how much was just the club's form.
- -Thousands of paths through the same match, which form the distribution - for example "68 out of 100 simulations".
- -The displayed result is the most common exact score. Minutes, scorers and the timeline come from a representative iteration that ended with that score.
What we do not count: weather, kilometres on the coach, derby temperature. The engine can do all of it, but those corrections are copied from the literature and we have not yet checked on our own settled matches whether they help at all. Until we do, they stay off. We would rather run a narrower model than a longer list.
The exact weights, the order of adjustments or the full feature list. That is the only thing we keep under the hood - and the only reason your result cannot be reproduced with a pen and a calculator.
See the model in action
Set a lineup and tactics to see how the model changes the match scenario.